Abstract
In the face of the escalating use of image editing tools and the rapid progress of AI technology, the demand for robust image forgery detection techniques has become more pressing. This study introduces a comprehensive approach to image forgery detection, which integrates image metadata analysis and Error Level Analysis (ELA) with deep learning, specifically Convolutional Neural Networks (CNNs) and Multilayer Perceptrons (MLPs). Using transfer learning, we evaluated seven pre-trained models alongside a custom innovative CNN model using the CASIAV2 dataset for binary classification of authentic and forged images. Metadata was extracted from all the images, and unique feature images were generated and used as inputs for all models by exploiting the differences between compressed and original images through ELA. Among the pre-trained models, VGG16, Xception, ResNet101, and MobileNetV2 demonstrated superior detection accuracy compared to the state-of-the-art models. However, our custom model surpassed six others, achieving an accuracy score of 99.08% with efficient training parameters. This study's results indicate significant advancements in image forgery detection, which has practical applications in forensic tools, automated detection systems, and educational purposes. An enhanced accuracy can benefit industries reliant on image authenticity, improve digital security protocols, protect intellectual property, and influence legal standards. Additionally, the methodology could extend to other domains like document and video authentication.